Conference Agenda

As the satellite altimetry data record is now nearly 30 years old, it is possible to use deep learning to predict sea-level changes based on observations. Using Temporal Fusion Transformers, we propose a global- and regional sea-level change Python-based prediction framework that includes multivariate inputs, explainability, and modeling uncertainty.
By training a network on TOPEX/Jason/Sentinel-6 sea-level anomaly time series from multiple grid points, we construct a regional model that can capture non-linear spatiotemporal relationships from which we derive statistics and compare them to current empirical and mathematical models.
Even though Global Mean Sea Level (GMSL) changes linearly with time (3 mm/year), this global average exhibits significant geographical variations. In addition, it covers a suite of regional non-linear signals varying in space and time.
Because deep learning can capture the non-linearity of the system, it offers an intriguing promise.
Furthermore, improving the mapping and understanding of these regional signals will enhance our ability to project sea level changes into the future.
This project focuses on the regional setting of the Western Pacific.

Harmful Algal Blooms (HABs) pose a great threat to human and animal health, their occurrence also has a significant impact on a variety of socio-economic and environmental factors. HAB events are now a global problem which affect food production, tourism, and ecosystem health. It is expected that the occurrence of HABs is likely to grow significantly with the increase in human population coupled with climate change. Phytoplankton size class (PSC) is suggested to be a good indicator of cell size, and considered to reflect the ecological and biogeochemical functional role of the phytoplankton present in the water column. Thus, it is important to be able to monitor PSCs, particularly in dynamic coastal waters where there are frequent changes in nutrients and phytoplankton community structure.

This study will draw on satellite sensors which differ in spatial, spectral, and temporal resolution: Sentinel-2 MSI, Sentinel-3 OLCI.The objectives of this study are to develop and validate HAB detection and Phytoplankton size classes (PSC) algorithms for near-shore and coastal waters, with better generalisation capability and lower computational overload that could improve the identification of the optical characteristics directly associated with phytoplankton properties

The research will be focused on four optically diverse regions of interest; The Danube Delta and Black Sea Coastline (Romania), Galician Coast (NW Spain), Shandong Peninsula Coast (China) and the Northern-South China Sea (China). Here, we will present results from the Galician coast and other European waters. We used in-situ data such as hyperspectral Remote Sensing Reflectance, Chlorophyll-a concentration, phytoplankton abundance and taxonomy, along with fractionated chlorophyll-a and particle absorption properties to develop and test the algorithms. We focus on the detection of Alexandrium minutum from Sentinel-2 MSI and Sentinel-3 OLCI data. Existing PSC retrieval algorithms based on pigment cover, chlorophyll-a abundance, and phytoplankton absorption for coastal and transitional waters were tested. In addition, We tested different atmospheric correction models against in-situ hyperspectral data and evaluated their performance over coastal waters.

We will present results on the optical characteristics of A. minutum and the potential of MSI and OLCI for their remote detection. We will discuss our plans for the development of Super Learners for HAB indicators and PSC and the evaluation of the PSC algorithms.

South China Sea (SCS) is one of the regions with frequent internal waves across the global ocean, making it the ideal bed for testing numerical models as well as the satellite measurements. In this study, we take advantage of the high-resolution hydrostatic MITgcm to constitute a dedicated model for the internal waves in SCS. The model outputs including three-dimensional current/temperature/salinity and surface height are at spatial resolution of 250 m every 3 hours. A practical algorithm to identify occurrence of internal waves events is proposed based on the gradient of surface height. Validations with the spaceborne synthetic aperture radar (SAR) measurements acquired during 2010-2019 as well as the optical MODIS observations show high consistency. This adds strong confidence to the performance of our configured MITgcm in future application of predicting and/or diagnosing the internal waves in SCS.

One of the most important applications of remote sensing (RS) technologies is about the detection and monitoring of ground surface changes exploiting using multi-temporal, remotely-sensed images [1–2]. In this framework, optical RS sensors have extensively been applied for change detection in a variety of heterogeneous applications. Essentially, change detection is a process that analyzes two or more images captured over the same geographical area at different times to identify those significant land cover changes that have occurred. Unlike optical sensors, microwave RS images acquired by synthetic aperture radar (SAR) have less been exploited for CD purposes. Despite its complexity, SAR images in change detection [3–4] are attractive from the operational viewpoint since SAR sensors are active instruments that operate in any atmospheric and sunlight conditions. In this work, we clarify the potential of coherent and incoherent CD approaches and we introduce the practical use of synthetic CD indices that can jointly be used to extract changed areas. In this context, artificial intelligence (AI) algorithms have already demonstrated their value [5-6]. Specifically, we generated series of CDI’s and we trained a random forest using these data and a set of external information on the state of observed scenes. The developed method has been tested considering as main events responsible for the observed changes the fires affecting the Sardinia and Sicily Island in 2021 as well as a flood event occurred in Houston Galvestone bay area. The main objectives of this study, whose results have recently published in [7], are to discuss the statistical properties, as well as the pros and cons, of CDIs based on the use of SAR backscattering (sigma nought) and InSAR coherence maps for the fast detection of changed areas. The role of machine learning methodologies, and the specific use of Random Forests (RF) [8] in CD tasks, is emphasized and some results obtained using Sentinel-1 SAR data are presented. Experimental results demonstrate the validity of the proposed integrated SAR/RF method for the fast mapping of fired/flooded zones and can further be extended to perform analyses in other contexts, such as the analysis of changes occurring in urban scenarios, considering the specific characteristics of these new environments.

References

1. Hansen, M.C.; Loveland, T.R. A Review of Large Area Monitoring of Land Cover Change Using Landsat Data. Remote Sens. Environ. 2012, 122, 66-74.
2. Bruzzone, L.; Prieto, D.F. Automatic Analysis of the Difference Image for Unsupervised Change Detection. IEEE Trans. Geosci. Remote Sens. 2000, 38, 1171–1182.
3. Bovolo, F.; Bruzzone, L. A Detail-Preserving Scale-Driven Approach to Change Detection in Multitemporal SAR Images. IEEE Trans. Geosci. Remote Sens. 2005, 43, 2963–2972.
4. Conradsen, K.; Nielsen, A.A.; Sehou, J.; Skriver, H. A Test Statistic in the Complex Wishart Distribution and Its Application to Change Detection in Polarimetric SAR Data. IEEE Trans. Geosci. Remote Sens. 2003, 41, 4–19.
5. Khelifi, L.; Mignotte, M. Deep Learning for Change Detection in Remote Sensing Images: Comprehensive Review and Meta- Analysis. IEEE Access 2020, 8, 126385–126400.
6. Shi, W.; Zhang, M.; Zhang, R.; Chen, S.; Zhan, Z. Change Detection Based on Artificial Intelligence: State-of-the-Art and Challenges. Remote Sens. 2020, 12, 1688.
7. Mastro, P.; Masiello, G.; Serio, C.; Pepe, A. Change Detection Techniques with Synthetic Aperture Radar Images: Experiments with Random Forests and Sentinel-1 Observations. Remote Sens. 2022, 14, 3323. https://doi.org/10.3390/rs14143323
8. Breiman,L.RandomForests.Mach.Learn.2001,45,5–32.

The measures of the Earth’s ground deformation by using multi-temporal interferometric synthetic aperture radar (InSAR) algorithms [1], [2] is nowadays a consolidated and mature practice that allows to have a really high accuracy, of the order of few millimeters [3], [4] in the line-of-sight direction of the SAR sensor. Among the different multi-temporal InSAR processors, a significant role is played by those algorithms based on the use of small-baseline (SB) multi-look (ML) interferograms [2], [5], which are less affected by decorrelation noise artefacts [6]. The conventional ML interferograms are independently generated by averaging adjacent neighbor pixels, in this way, the signal-to-noise ratio drastically increases and the analysis of the distributed scatterers becomes possible, or to some extent, less challenging. The previous multi-look operation involves the average of the information relating to each family of scatterers present in the single look pixels that will contribute to each multi-look pixel. Recently, in [7] has been observed that some inconsistencies in the InSAR products (i.e., ground deformation time-series and mean deformation velocity maps) may happen when sets of multi-look SAR interferograms with very short temporal baselines are processed, compared to those obtained using interferograms with longer temporal baselines. Such spurious signals lead to systematic biases [7] that, if not adequately compensated for, might lead to unreliable InSAR ground displacement products.

In this work, we propose a technique to estimate and correct a set of multi-look SB interferograms that is based on computing and analyzing exclusively sets of (wrapped) non-closure phase triplets. The developed phase estimation method works on every single SAR pixel independently, assuming the (unknown) phase bias signal could be approximated with a second order expansion, basically, as the sum of a constant phase velocity term v and a time-dependent (i.e., dependent on the interferograms temporal baseline) phase velocity difference terms Δv(Δt_i), where Δt_i is the temporal baseline of the generic i-th interferogram. Once the whole set of triplets that could be formed using short baseline ML interferograms is identified, and considering the mathematical properties of the triplets non-closure phases, we can write an overdetermined system of linear equations, where the known terms are the measured wrapped non-closure phases over the set of identified triplets, namely ΔΦ_triplets, and the unknowns are the temporal-baseline-dependent phase velocity difference terms Δv. For example, considering the Sentinel 1-A/B sensors, the temporal baseline is sampled with an atomic sampling time of six days; accordingly, if we accept, for instance, a threshold of 96 days for the maximum allowed temporal baseline of the selected SB interferograms, we have 16 unknowns to be estimated. Once the linear system is solved in the Least-Squares sense, the phase biases at the different temporal baselines, namely ΔΦ_bias, are iteratively retrieved by doing a backward integration step, assuming as the initial condition that the phase bias at the maximum considered temporal baseline is zero, that is Δφ_{bias(max_temp_bas)}=0.

Simulated tests show the validity of the proposed method. In addition, real experiments performed on sets of Sentinel1-A/B SAR data in different geo-morphological conditions, are in accordance with the synthetic experiment, confirming the effectiveness of the developed methodology. Furthermore, we performed some simulations and experiments to test the validity of an extension of the developed method to the non-stationary case, e.g., when the phase bias signals depend on the specific single time acquisitions of the SAR images involved in the SB interferograms generation, and not only on their temporal baselines. Further developments are still needed to understand what useful information we can extract from these estimated phases that represent a bias on displacement estimation, but they could be a resource for estimating other signals, such as being soil moisture.

References:

[1] A. Ferretti, C. Prati, and F. Rocca, «Permanent scatterers in SAR interferometry», IEEE Trans. Geosci. Remote Sens., vol. 39, n. 1, pagg. 8–20, gen. 2001, doi: 10.1109/36.898661.

[2] P. Berardino, G. Fornaro, R. Lanari, and E. Sansosti, «A new algorithm for surface deformation monitoring based on small baseline differential SAR interferograms», IEEE Trans. Geosci. Remote Sens., vol. 40, n. 11, pagg. 2375–2383, 2002.

[3] A. Ferretti et al., «Submillimeter Accuracy of InSAR Time Series: Experimental Validation», IEEE Trans. Geosci. Remote Sens., vol. 45, n. 5, pagg. 1142–1153, mag. 2007, doi: 10.1109/TGRS.2007.894440.

[4] F. Casu, M. Manzo, and R. Lanari, «A quantitative assessment of the SBAS algorithm performance for surface deformation retrieval from DInSAR data», Remote Sens. Environ., vol. 102, n. 3, pagg. 195–210, giu. 2006, doi: 10.1016/j.rse.2006.01.023.

[5] F. Falabella, C. Serio, G. Zeni, and A. Pepe, «On the Use of Weighted Least-Squares Approaches for Differential Interferometric SAR Analyses: The Weighted Adaptive Variable-lEngth (WAVE) Technique», Sensors, vol. 20, n. 4, Art. n. 4, gen. 2020, doi: 10.3390/s20041103.

[6] H. A. Zebker and J. Villasenor, «Decorrelation in interferometric radar echoes», IEEE Trans. Geosci. Remote Sens., vol. 30, n. 5, pagg. 950–959, set. 1992, doi: 10.1109/36.175330.

[7] H. Ansari, F. De Zan, and A. Parizzi, «Study of Systematic Bias in Measuring Surface Deformation With SAR Interferometry», IEEE Trans. Geosci. Remote Sens., vol. 59, n. 2, pagg. 1285–1301, feb. 2021, doi: 10.1109/TGRS.2020.3003421.

In the context of global climate change, frequent natural disasters have caused great harm to economic and social development, human life safety, ecosystems and many other aspects, and have a profound impact on regional sustainable development. According to the EM-DAT database (Emergency Events Database), there were about 4,200 major natural disasters in the world with more than 1 million deaths, 2.1 billion people directly affected, and direct economic losses as high as $1.7242 trillion from 2004 to 2014.

As the economic center of China, Shanghai is a mega-city with a permanent population of more than 24 million. There are dense various kinds of buildings and major infrastructures. As a coastal city, it is vulnerable to typhoon, rainstorm, waterlogging, storm surge and red tide. Besides, it is also under threat of lightning strike and high temperature due to global warming and its location in the monsoon region. According to the historical disaster survey of the First National Survey on Integrated Risk of Natural Disasters in Shanghai, the natural disasters that have a greater impact on Shanghai are mainly typhoon, flood and hail. Coastal districts in Shanghai, including Pudong New District, Baoshan District, Fengxian District, Jinshan District and Chongming District, are also affected by marine disaster such as storm surges. In addition, Shanghai is located at the mouth of the Yangtze River, formed by alluvial sediment. As a result, geological disasters such as land subsidence are also prone to occur with the rapid economic development and frequent land reclamation activities. Therefore, it is of great significance to carry out assessment of natural disaster mitigation capability, and then to improve natural disaster mitigation capability. Based on this, the government can make disaster-prevention strategies to minimize casualties and property damage when disaster strikes.

This study is based on the index system of natural disaster mitigation capability in the First National Survey on Integrated Risk of Natural Disasters. TOPSIS（Technique for Order Preference by Similarity to an Ideal Solution）is utilized to assess natural disaster mitigation capability in Xuhui District at township level as assessment unit. The capability of disaster reduction of the 13 streets and towns in Xuhui district are divided into four levels, "strong", "secondary strong", "medium" and "weak". We analyzed the impact of weight change of the index system on the assessment results. The results show that weight change can change the assessment results, but the changes are not significant. Furthermore, we found out the key index that affect the assessment on natural disaster mitigation capability on township level in Xuhui District by simulating various reasonable values and made thematic maps of natural disaster mitigation capability of Xuhui District.

Keywords: the First National Survey on Integrated Risk of Natural Disasters; assessment of natural disaster mitigation capability; TOPSIS; weight change; key index; Xuhui District; Shanghai

Chongming Island is the third largest island in China. It is also the largest estuarine alluvial island in the world. The island is located in the mouth of the Yangtze River, covering an area of 1361.11 km². The shallow strata in Chongming Island are formed late, and most of them are marine continental interactive sedimentation. The large void ratio, high water content and incomplete natural consolidation process of soil layer lead to significant compression deformation. Statistic datasets show that the maximum land subsidence was more than –500 mm, and the subsidence rate exceeded –50 mm/year from 1995-2005 in the reclamation area of eastern Chongming Island. Due to the long land formation in the west of Chongming Island, the subsidence rate also reached –5.0 mm/year in the same period. Serious land subsidence can not only damage public infrastructure and buildings, but also lead to natural disasters such as seawater backflow and storm surge. Therefore, it is necessary to monitor the settlement of Chongming Island. The objective of this study is to investigate the ground deformation in Chongming Island by using three space-borne Synthetic Aperture Radar (SAR) satellite datasets to retrieve the recent decade deformation time-series of Chongming Island, and is to address the problem of evaluating the error budget of multi-platform MT-InSAR deformation time-series combination.

Three independent SAR datasets (from Radarsat-2 Multi-Look Fine (RST-2MF), Radarsat-2 Wide (RST-2W) and Sentinel-1A (S1A)) were used to investigate the land subsidence in Chongming Island. The first dataset consisted of 21 SAR images, acquired by the Radarsat-2 sensors in the multi-look fine (MF) mode from August 2009 to December 2013. The second SAR dataset was collected acquired by the Radarsat-2 sensors in the wide mode, from January 2012 to December 2016 and consisted of 44 SAR images. The third dataset consisted of 120 SAR images, acquired by the S1A sensors from July 2015 to June 2020. The Small Baseline Subset (SBAS) algorithm is independently applied to available SAR datasets to generate the corresponding ground displacement time-series. Then, the singular value decomposition (SVD) method is used to jointly analyze the deformation time series of datasets with time overlap. Finally, the accuracy of joint time series obtained by SVD algorithm is analyzed by using the error propagation law. First, ground deformation time-series of the three SAR satellite datasets were retrieved by SBAS. On this basis, the SVD method is used to jointly analyze the deformation time series of the common high coherence points of RST-2MF and RST-2W, so as to obtain the RST-2MF/RST-2W-SBAS deformation time series with a time span from August 2009 to December 2016. Then, the RST-2MF/RST-2W-SBAS deformation time series and S1A-SBAS deformation time series are jointly analyzed with the same method to obtain the RST-2MF/RST-2W/S1A-SBAS deformation time series from August 2009 to June 2020. Finally, aiming at the noise caused by decoherence as the source error of SVD method, the error propagation theory is applied to all common high coherence points in the study area, so as to obtain the error distribution of SVD method in Chongming Island.

Deformation time series derived from SBAS retrieved using 2009-2013 RST-2MF, 2012-2016 RST-2W and 2015-2020 S1A datasets. To combine RST-2MF-SBAS, RST-2W-SBAS and S1A-SBAS deformation time series in Chongming Island, a long-term deformation time series more than 10 years (2009-2020) is obtained. The statistical analysis results show that the mean values of deformation accumulation in three period (2009-2013，2012-2016 and 2015-2020) are, –16 mm, –11 mm and –7 mm respectively, and the corresponding annual average deformation rates are –5 mm/year, –1.7 mm/ year and –1.5 mm/year respectively. The RST-2MF/RST-2W/S1A-SBAS deformation time series shows that 98% of high coherence points of the cumulative deformation variables in the study area in recent decade are distributed between –120 mm ~ 20 mm, and the corresponding annual average deformation rate is distributed between –15 mm/year ~ 2 mm/year. The settlement rate of some reclamation areas in the east of Chongming exceeds –15 mm/year, and the cumulative settlement in recent decade has reached –200 mm. The error distribution of Chongming Island SVD method with a mean value of 5 mm and a standard deviation of 4 mm is obtained by error propagation method.

Change detection is to quantitatively analyze and determine the characteristics and process of earth surface change based on remote sensing data in different periods. It is widely used in disaster dynamic detection, urban planning and other fields[1]. With the impact of globalization, monitoring of urban changes with remote sensing images is becoming more and more important. Among remote sensing images, synthetic aperture radar (SAR) images are independent of solar illumination and atmospheric condition[2]. In this study, we focus on SAR images that are the sources of useful information in change detection. Speckle noise is an inherent problem of SAR images. It greatly influences the performance of change detection using SAR images. The study proposes a change detection method based on an end-to-end UNet with residual blocks(Res-UNet) model, in which the residual blocks can avoid the gradient disappearance and gradient explosion caused by the deepening of the network[3], and the encoding and decoding structure can improve the robustness to different levels of noise.

This study uses high-resolution TerraSAR-X(TSX) time-series images covering Shanghai. The acquisition time of TSX datasets is from 16 October 2015 to 19 August 2016. Before training the network, we pre-process TSX images to get amplitude information and coherent information. Amplitude information is obtained by pre-processing single look complex (SLC) images, including radiometric correction, multilooking processing and geocoding. Coherent information is obtained by performing interferometry and computing coherence values of interferograms. Res-UNet model is utilized to distinguish between changed and unchanged areas, which mainly consists of an encoder and an decoder. The encoder composed of four sequential convolution blocks with residual structure, each followed by a max pooling operation, is used to extract change features from the input images. The decoder is utilized to restore the input to its original size, which is composed of four sequential blocks of a transpose convolution operation with a stride of 2, each followed by the previously described convolution blocks. Res-UNet also uses skip connections between the encoder and the decoder, adding fine-grained shallow features to coarse-grained deep features[4]. Finally, a 1 × 1 convolution followed by a sigmoid activation function is used to create a one-band output with values between 0 and 1.

We evaluate the accuracy of Res-UNet model using amplitude information and coherent information. The results show that the method is reliable in urban change detection and performs better than other advanced change detection methods. Its Precision, Recall, F1-Score and Overall Accuracy reach 84.3%, 76.6%, 80.2% and 99.7% respectively.

The maximum wave height (Hmax) is an extremely important factor that has a significant impact on the safety of maritime activities. However, the Hmax is much less investigated than significant wave height (SWH) in the wave remote sensing. Nowadays, radar altimeters and CFOSAT provide the SWH operational but without Hmax products. A method of obtaining the Hmax from CFOSAT SWIM Level 2 parameters is presented. The buoys are the most reliable way to observe the Hmax, but the collocations between buoys and CFOSAT tracks are too few to perform the supervised learning training. The ERA5 wave reanalysis from the European Centre for Medium-Range Weather Forecasts (ECMWF) is one of the most accurate datasets. However, the obvious bias and scatter index of Hmax are found from the comparison between ERA5 and buoys located west of France. A machine learning model is firstly built to reduce the error of ERA5 Hmax. Then the corrected ERA5 Hmax is collocated with CFOSAT observations and used for the training target of SWIM Hmax retrieval. The SWIM parameters both from SWIM nadir and boxes, including the SWH, wavelength and wave partition information, are used to obtain the Hmax based on machine learning. The CFOSAT data in 2021 are used to train the Hmax machine learning model while the data in 2020 are used to perform the independent validation. The bias, RMSE and scatter index of CFOSAT Hmax are 0.01m, 0.51m, 16%, while 0.77m, 1.09m, 19% are for the ERA5. Therefore, this study provides a perspective to obtain the Hmax from satellite remote sensing for further applications such as marine forecasts.